About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Frontiers in Solidification X
|
| Presentation Title |
Physics-Informed Neural Operator for Phase Field Models of Solidification |
| Author(s) |
Chih-Kang Huang, Ludovick Gagnon, Benoît Appolaire, Miha Zaloznik |
| On-Site Speaker (Planned) |
Miha Zaloznik |
| Abstract Scope |
The long computation times for the phase-field simulation of microstructures could be overcome with surrogate AI models based on neural networks.
We propose a new neural-operator approach for phase-field models of solidification. It consists of a Deep Ritz method, where a neural network is trained to approximate a variational formulation of the phase-field model. The training is unsupervised and relies exclusively on satisfying the model equations. We further introduce a custom Reaction-Diffusion Neural Operator (RDNO) architecture, adapted to the structure of the model equations.
We successfully apply the neural-operator approach to the Allen-Cahn equation and to dendritic growth simulation. We demonstrate that our physically-informed training provides better generalization in out-of-distribution evaluations than data-driven approaches. We also show that the RDNO architecture outperforms popular standard architectures (MLP, FNO, and U-Net) in our test cases. Finally, we achieve evaluation speeds about four times faster than conventional Fourier spectral schemes. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Solidification, Modeling and Simulation, Machine Learning |